Fig. 1. Overall structure of assessment model
Fig. 2. Single-channel assessment framework based on parallel CNN
Fig. 3. VGG16 network structure
Fig. 4. Influences of blur and noise on visual characteristics of remote sensing images. (a) Original image; (b) image with noise; (c) image with noise once more; (d) image with blur; (e) image with blur once more
Fig. 5. Remote sensing images with different texture complexity. (a) S=9.2784; (b) S=23.1248; (c) S=19.0502; (d) S=14.9255; (e) S=20.1074; (f) S=13.2125; (g) S=23.2592; (h) S=18.7957
Fig. 6. Remote sensing images acquired by QuickBird-2 satellite. (a) Harbor; (b) vegetation; (c) road; (d) buildings
Fig. 7. Overall assessment results by proposed method
Fig. 8. Fitting scatter plot between proposed method and SSIM, PSNR, FSIM. (a) SSIM; (b) PSNR; (c) FSIM
Fig. 9. Fitting curves of DMOS for different assessment methods in LIVEMD dataset. (a) SSIM; (b) proposed method; (c) PSNR; (d) FSIM
i | 1 | 2 | 3 | 4 | 5 |
---|
Bblur_i | 5 | 4 | 3 | 2 | 1 | Blurintensity | 0 | 1 | 2 | 3 | 4 |
|
Table 1. Blur level
j | 1 | 2 | 3 | 4 | 5 |
---|
Nnoise_j | 5 | 4 | 3 | 2 | 1 | Noiseintensity | 0 | 1 | 2 | 3 | 4 |
|
Table 2. Noise level
Five-grade scale | 5 | 4 | 3 | 2 | 1 |
---|
Quality | Excellent | Good | Fair | Poor | Bad | Impairment | Imperceptible | Perceptible,but not annoying | Slightlyannoying | Annoying | Veryannoying |
|
Table 3. ITU-R quality and impairment scales
Method | PLCC | RMSE | SROCC |
---|
SSIM | 0.8979 | 0.5631 | 0.8714 | PSNR | 0.8450 | 0.6842 | 0.7713 | FSIM | 0.8978 | 0.5602 | 0.8793 | Proposed | 0.9004 | 0.5566 | 0.8962 |
|
Table 4. Performance comparison of different methods
Method | PLCC | RMSE | SROCC |
---|
SSIM | 0.7679 | 11.9487 | -0.6953 | PSNR | 0.7757 | 11.7910 | -0.7088 | FSIM | 0.8178 | 10.7358 | -0.8642 | Proposed | 0.8968 | 8.2523 | -0.8664 |
|
Table 5. Performance comparison of different methods in LIVEMD database